Detection is nearly impossible. Prevention is the only real option.
In classrooms across Spain, a quiet crisis is unfolding: the tools built to catch deception cannot keep pace with the tools built to create it. As artificial intelligence makes it trivially easy to generate convincing academic work, educators are discovering that the answer lies not in surveillance but in reimagining what it means to truly know something. The question is no longer whether a student used a machine, but whether the systems we built to measure learning were ever measuring the right things.
- Detection tools for AI-generated work are fundamentally broken — unlike plagiarism software, no reliable system exists, and students have already learned to exploit the gaps by deliberately introducing errors and simplifying language.
- The divide in the classroom is not just technological but motivational: some students chase genuine understanding while others use AI purely to survive the system, and both sit in the same room under the same rules.
- Teachers like Ana Eva Santos in Valencia have stopped chasing fraud after the fact and instead redesigned the conditions of learning itself — banning homework, locking away devices, and treating AI as a teaching instrument rather than a threat.
- Institutions from Spanish secondary schools to Wharton Business School are converging on the same solution: oral exams, continuous evaluation, and practical exercises that force students to demonstrate knowledge in ways no algorithm can easily fake.
- The real transformation is not in detection software but in the questions educators are willing to ask — and whether the systems built to measure learning were ever truly measuring learning at all.
Ana Eva Santos moves through her English classroom at Valencia's Pío XII school with a skill no software has mastered: she knows her students well enough to recognize when the work is not theirs. Her response was not to hunt for fraud but to eliminate the conditions that make it easy. She banned homework entirely. Every assignment happens in her classroom, under her watch, with devices locked away or placed visibly on desks. She uses AI as a teaching tool, not a target.
Elsewhere, the problem runs deeper. At Madrid's Ramón y Cajal school, students have learned to make AI-generated text look human — simpler sentences, deliberate spelling errors, limited vocabulary. These are not students chasing excellence; they are students trying to pass. Two entirely different motivations share the same classroom.
The fundamental difficulty is that no reliable detection tool exists for AI-generated work, unlike the plagiarism detectors educators once trusted. José García Montalvo, an economics professor at Universitat Pompeu Fabra, is candid: detection is nearly impossible, and prevention is the only real option. This might mean restricting students to a single provided database, or banning assignments that AI can complete in seconds. Experimental watermarking technologies — analogous to security features on currency — are being developed, but remain unproven at scale.
Existing detection methods rely on statistical patterns: unusual consistency in sentence length, rigid vocabulary, structural repetition. Remote proctoring software, frequency detectors, and stricter in-person controls are spreading across Spain, though enforcement varies sharply by region. Some autonomous communities have implemented these measures for university entrance exams; others have not.
A more promising shift is underway. Digital consultant Miguel Ángel Corella observes that institutions are moving toward oral exams, practical exercises, and continuous evaluation — formats that make automated answers far harder and force a genuine reckoning with what it means to demonstrate knowledge. Even Wharton Business School has followed this logic, banning computers from in-class exams and distinguishing between graduate students permitted to use AI and undergraduates who must first learn to think and write on their own.
At KPMG, hiring now favors in-person interviews, and thousands of employees have completed extensive AI training. The arms race between detection and evasion continues — but the more consequential transformation is happening in how schools ask questions, how they evaluate answers, and what they ultimately want to know.
Ana Eva Santos walks between the desks in her English classroom at Pío XII, a school in Valencia, and she has learned something that no detection software can quite teach: how to spot a lie by knowing her students. After years of teaching, she recognizes the moment when a struggling pupil suddenly submits work that reads like it came from an Oxford scholar. The answer, she discovered, was not to hunt for the fraud after the fact but to prevent it from the start. She banned homework entirely. Every assignment happens in her classroom, during her watch, with phones and smartwatches either locked away or visible on the desk. She uses artificial intelligence as a teaching tool, not as something to police.
But the problem is real, and it runs deeper than one teacher's classroom rules. At Ramón y Cajal school in Madrid, Héctor García, a strong student himself, observes that his peers have learned to game the system in subtler ways. They use AI tools that deliberately humanize the output—simpler language, occasional spelling mistakes, basic sentence structures, limited vocabulary, repetitive phrasing. These are not the students chasing excellence. They are the ones simply trying to pass, to move through the system without distinction. Two entirely different motivations sit in the same classroom: one seeking genuine learning, the other seeking mere survival.
The core problem is this: there is no foolproof tool to detect AI-generated work, unlike the plagiarism detectors of the past. José García Montalvo, an economics professor at Universitat Pompeu Fabra in Barcelona, speaks with candor about the difficulty of finding deception in his field. Detection, he says, is nearly impossible. Prevention is the only real option. This might mean requiring students to work from a single, provided database. It might mean banning business plans that are simply repackaged versions of existing strategies. With ChatGPT and Claude, such shortcuts are trivially easy to create.
The impossibility of reliable detection is forcing schools to rethink what they teach and how they teach it. Montalvo notes that the inability to catch AI use is reshaping the very questions professors ask their students. Tools like GPTZero, AI Detector, Copyleaks, and QuillBot exist, but each has fractures—weak points where a simple tool can break through. Some experimental projects are attempting to create a kind of watermark, similar to security features on currency, that would distinguish AI-generated content from original work. But these remain in early stages, unproven at scale.
The detection methods that do exist rely on statistical patterns: in text, looking for unusual consistency in sentence length and word frequency, rigid vocabulary choices, and structural repetition; in images, examining reflections and textures; in voice, using machine learning classifiers to distinguish synthetic speech from genuine. Schools across Spain and globally have begun reinforcing their monitoring systems. Remote proctoring software like Smowl watches students during online exams. In-person tests now feature stricter physical controls—no electronic devices, frequency detectors to catch hidden communication tools like earpieces or smartwatches. But enforcement varies wildly by region, with some autonomous communities implementing these measures for university entrance exams this year and others holding back.
A different approach is gaining ground, one that sidesteps the detection problem altogether. Miguel Ángel Corella, a digital solutions consultant, observes that institutions are shifting toward oral exams, practical exercises, and continuous evaluation. These formats make it far harder to automate answers and force a deeper reckoning with what it actually means to demonstrate knowledge in an age of AI. Even Wharton Business School, one of the world's most prestigious institutions, has adopted this logic. Computers are banned from in-class exams, and software monitors for AI use. Graduate students are permitted to use AI; undergraduates are not, because they must first learn to write and think clearly on their own. Perhaps the solution is simpler than any algorithm: can artificial intelligence really write something like this in twenty-six words? "All is clouded by desire: as fire by smoke, as a mirror by dust. Through them the soul is blinded." The Year of Living Dangerously, Peter Weir, 1982.
At KPMG, one of the world's largest consulting firms, the hiring process now emphasizes in-person interviews over remote ones, where AI tools complicate genuine conversation. The company's 6,430 Spanish employees have completed 140,000 hours of training in this technology. They have learned that all detection systems work the same way: analyzing sentence length, word frequency, structural consistency, vocabulary rigidity, paragraph uniformity, and comparison to a student's previous writing style. Yet they all fail sometimes. The arms race between detection and evasion continues, but the real shift is happening elsewhere—in how schools ask questions, how they evaluate, and what they actually want to know.
Notable Quotes
A student who has struggled all semester and suddenly submits work that reads like an Oxford academic—you know there is deception.— Ana Eva Santos, English teacher at Pío XII school, Valencia
Only prevention is possible. Detection of AI use is nearly impossible in my field.— José García Montalvo, economics professor at Universitat Pompeu Fabra, Barcelona
The Hearth Conversation Another angle on the story
Why can't we just build a better detector? We have the technology.
We do, but detection is always playing catch-up. Every tool has weaknesses. A student can prompt an AI to write badly on purpose—add typos, use simple words, break up sentences. The detector looks for patterns, but humans are pattern-breakers.
So you're saying it's impossible?
Not impossible. Just not solvable through detection alone. That's why the smart schools are changing the game entirely—they're asking questions that can't be answered by a machine, or at least not easily.
Like what?
Oral exams. A conversation where you have to think on your feet. Practical work you do in front of someone. Continuous evaluation over time, so they know your actual voice. You can't fake that.
But that's more work for teachers.
Yes. But it's also more honest. It actually tests whether you understand something, not whether you can prompt an AI well.
What about the students who just want to pass?
They still have to show up. They still have to think. The system stops rewarding the shortcut.